memos

MemOS is a memory operating system for LLMs and AI agents. It unifies store, retrieve and manage operations for long-term memory, aiming at context-aware and personalized interactions, with knowledge bases, multi-modal data, tool memory and enterprise-oriented optimizations built in. Key capabilities described in the README: - Unified Memory API: one API to add, retrieve, edit and delete memory, structured as a graph that is inspectable and editable rather than an opaque embedding store. - Multi-modal memory: text, images, tool traces and personas are supported and can be retrieved and reasoned over together. - Multi-cube knowledge base management: multiple knowledge bases act as composable memory cubes, allowing isolation, controlled sharing and dynamic composition across users, projects and agents. - Asynchronous ingestion via MemScheduler: memory operations can run asynchronously with millisecond-level latency for stability under high concurrency. - Memory feedback and correction: memories can be corrected, supplemented or replaced through natural-language feedback. Deployment options: - Cloud API: a hosted service accessed with an API key from the MemOS dashboard; add and search memories over REST endpoints. - Self-hosting: Docker Compose starts the MemOS API together with Neo4j and Qdrant, or the API can be run directly with uvicorn if those services are already available. Configuration covers LLM provider, embedder, vector DB, graph DB and scheduler. - Plugins: a cloud plugin and a local plugin integrate memory into OpenClaw, Hermes Agent and DeepSeek Harness. The cloud plugin recalls memories before an agent run and saves new messages afterwards, and is described as fail-open so a temporary outage does not interrupt the task. The local plugin stores memory on-device in SQLite and provides hybrid retrieval combining FTS5 and vector search, smart deduplication, tiered skill evolution across L1 traces, L2 policies and L3 world models, multi-agent collaboration, and a Memory Viewer dashboard. Installation paths include an npm package for the cloud plugin, shell and PowerShell installers for the local plugin, and Docker or uvicorn for the self-hosted service. The project is licensed under Apache 2.0 and links to documentation, an arXiv paper, Discord and GitHub discussions. The README also reports benchmark scores and token-savings figures; these are the project's own claims and are not independently verified here.